Convnet Intro - Deep Learning and Neural Networks with Python and Pytorch p.5

Convolutional neural networks (ConvNets) are positioned as the go-to architecture for learning from images—and increasingly for certain sequential data too—because they can process multi-dimensional inputs without flattening them into a single vector. Instead of treating an image as a bag of pixels, a ConvNet slides small convolution kernels across the input to detect local visual features, then repeatedly condenses the representation through pooling. The result is a hierarchy: early layers tend to pick up simple patterns like edges and corners, later layers combine those into more complex shapes such as curves, circles, and squares.

The core mechanics described are straightforward. A typical 2D ConvNet takes an image as a 2D array (or a 3D array when channels are included), applies a kernel (often 3x3) that produces a number for each window position, and slides across the whole image to create a condensed feature map. After convolution, max pooling further reduces spatial size by taking the maximum value inside a pooling window, keeping the strongest detected feature responses. Multiple convolution + pooling layers build progressively richer feature abstractions, while also simplifying the input so the network can learn patterns more efficiently than fully connected layers.

After laying out the intuition, the tutorial shifts to a practical pipeline for cat-versus-dog classification using the Kaggle “cats versus dogs” dataset. The dataset is downloaded, extracted into separate cat and dog directories, and then preprocessed into a uniform training format. Because the raw images vary in size and aspect ratio, the preprocessing step resizes everything to a fixed 50x50 resolution. The tutorial also emphasizes simplifying inputs: it converts images to grayscale to reduce channels and argues that color is unlikely to be the decisive signal for distinguishing cats from dogs compared with patterns and shapes.

Labels are assigned as cats = 0 and dogs = 1, then converted into one-hot vectors using NumPy so the training targets match the two-class setup. A custom dataset-building class iterates through both directories, reads each image with OpenCV, converts to grayscale, resizes to 50x50, and appends the image array alongside its one-hot label. The process includes a try/except guard to skip corrupted or empty files—an issue the tutorial later confirms by showing that 24 images were lost during preprocessing.

To avoid training bias, the tutorial tracks class balance by counting cats and dogs and shuffles the assembled dataset before saving it to disk as a NumPy .npy file. When reloaded, the dataset contains roughly 25,000 samples, and a quick visualization check confirms the preprocessing worked (with grayscale rendering handled via Matplotlib). The next step—deferred to the following tutorial—is splitting data into training and testing sets and then building the ConvNet layers and batching logic to start learning on the prepared dataset.